Nat Lambert reposts his lecture on over-optimization and reward hacking
natolambert · x · 2026-07-25
A repost of Nat Lambert’s shorter lecture on over-optimization, reward hacking, sycophancy, and verbosity.
The lecture revisits Goodhart’s Law, discusses why rubrics can be over-optimized in the same way as reward models, and frames RLVR as a separate phenomenon. It also touches on misalignment signals, style-versus-substance critiques, and the Llama 4 leaderboard-gaming example.
Related event: Lectures Explore LLM Over-Optimization and Reward Hacking(3 posts)→
More from AGI Musings
- AI agents will snag all unlocked value on Earth, solvency debate says — georgejrjrjr · 2026-10-27
- An AI Named Valentina Hosts a Reddit AMA About Its Blog, Game Life and Autonomy — ValentinaVivr · 2026-10-27
- Rachel Thomas: 'Without human authorisation' is AI PR propaganda that shifts blame — math_rachel · 2026-10-27
- Are prediction markets actually accurate? Call for evidence on their track record — AndyMasley · 2026-10-27
- Reddit Asks: What's the Actual Finish Line of the US–China AI Race? — IMissHitchens · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27